the dumb zone: after ~40% of context used, it starts diminishing returns; YMMV
subagents are for managing context, not for specialized roles (QA, Architect, Data Scientist, etc.)
Amount of lies graph: the actual code < function names < comments < documentation
Frequent Intentional Compaction
research: compression of truth (real code, not docs); agents scanning the codebase and generating research.md (300-1000 lines);
human in the loop: read/review/fix
plan: compression of intent; outline exact implementation steps, including filenames, lines and snippets; explicit about testing steps (auto and manual); input: research.md`` + ticket; output: plan.md` (300-1000 lines); you must read/review the plan.